Trim insignificant weights

This document provides an overview on model pruning to help you determine how it fits with your use case.

• To dive right into an end-to-end example, see the Pruning with Keras example.
• To quickly find the APIs you need for your use case, see the pruning comprehensive guide.
• To explore the application of pruning for on-device inference, see the Pruning for on-device inference with XNNPACK.
• To see an example of structural pruning, run the tutorial Structural pruning with 2 by 4 sparsity.

Overview

Magnitude-based weight pruning gradually zeroes out model weights during the training process to achieve model sparsity. Sparse models are easier to compress, and we can skip the zeroes during inference for latency improvements.

This technique brings improvements via model compression. In the future, framework support for this technique will provide latency improvements. We've seen up to 6x improvements in model compression with minimal loss of accuracy.

The technique is being evaluated in various speech applications, such as speech recognition and text-to-speech, and has been experimented on across various vision and translation models.

API Compatibility Matrix

Users can apply pruning with the following APIs:

• Model building: keras with only Sequential and Functional models
• TensorFlow versions: TF 1.x for versions 1.14+ and 2.x.
  • tf.compat.v1 with a TF 2.X package and tf.compat.v2 with a TF 1.X package are not supported.
• TensorFlow execution mode: both graph and eager
• Distributed training: tf.distribute with only graph execution

It is on our roadmap to add support in the following areas:

• Minimal Subclassed model support
• Framework support for latency improvements

Results

Image Classification

Model	Non-sparse Top-1 Accuracy	Random Sparse Accuracy	Random Sparsity	Structured Sparse Accuracy	Structured Sparsity
InceptionV3	78.1%	78.0%	50%	75.8%	2 by 4
		76.1%	75%
		74.6%	87.5%
MobilenetV1 224	71.04%	70.84%	50%	67.35%	2 by 4
MobilenetV2 224	71.77%	69.64%	50%	66.75%	2 by 4

The models were tested on Imagenet.

Translation

Model	Non-sparse BLEU	Sparse BLEU	Sparsity
GNMT EN-DE	26.77	26.86	80%
		26.52	85%
		26.19	90%
GNMT DE-EN	29.47	29.50	80%
		29.24	85%
		28.81	90%

The models use WMT16 German and English dataset with news-test2013 as the dev set and news-test2015 as the test set.

Keyword spotting model

DS-CNN-L is a keyword spotting model created for edge devices. It can be found in ARM software's examples repository.

Model	Non-sparse Accuracy	Structured Sparse Accuracy (2 by 4 pattern)	Random Sparse Accuracy (target sparsity 50%)
DS-CNN-L	95.23	94.33	94.84

Examples

In addition to the Prune with Keras tutorial, see the following examples:

• Train a CNN model on the MNIST handwritten digit classification task with pruning: code
• Train a LSTM on the IMDB sentiment classification task with pruning: code

For background, see To prune, or not to prune: exploring the efficacy of pruning for model compression [paper].